Pre-processing and Hybridized Segmentation Strategies from SPIR based Liver Images Utilizing the Concepts of Machine Learning for Improvement of Enhanced Performances of Evaluation Parameters with Feature Extraction

Abstract

Biomedical imaging modalities such as Spectral Pre-saturation with Inversion Recovery (SPIR) MRI frequently suffer from non-Gaussian Rician noise, spatial field inhomogeneities, and weak parenchymal boundaries. This paper presents a complete, automated computer-aided diagnostic (CAD) pipeline designed for real-time edge processing and high-accuracy classification of liver pathologies. The proposed framework introduces a two-stage pre-processing pipeline combining non-linear bilateral spatial filtering with dynamic Luminance-Level Modulation (LM) and CLAHE to restore micro-texture contrast while suppressing artifacts. A Reaction-Diffusion Level Set Segmentation (RDLSS) model is implemented to eliminate periodic distance re-initialization and prevent boundary leakage across ill-defined lesion contours. A multi-domain spatial and spectral feature extraction scheme unifies 2D Discrete Wavelet Transforms (2D-DWT: Daubechies  / Symlets ) with Gray Level Co-occurrence Matrices (GLCM), Local Binary Patterns (LBP), hybrid Local Binarized GLCM (LBG-LCM), and Gray Level Run Length Matrices (GLRLM). An energy-ranked sub-band selection strategy reduces feature extraction overhead by up to 50%. The optimized feature vectors are classified using a Backpropagation Multilayer Perceptron (BP-MLP) network, achieving 96.00% classification accuracy, 95.80% sensitivity, 96.20% specificity, and a Dice similarity coefficient exceeding 0.94. Furthermore, a pipelined, fixed-point FPGA hardware co-processor architecture is presented to support real-time clinical deployment.